SDXL LoRA/LyCORIS works best on its original training model. You may need to reduce weight when using it with other models or additional LoRAs.
LyCORIS Combination Recommendation: Try using four of my LyCORIS models with the following weights: 0.4, 0.3, 0.2, 0.2 (or any combination with a total weight around 1.0).
I will release multiple versions for a few popular models. Feel free to request for artist or model.
Version naming convention: MajorVersion.MinorVersion [BaseModel]
MajorVesion: For big updates that can apply to any style LoRA. For example, training parameter update.
MinorVersion: For small updates that only apply to this LoRA. For example, epoch selection, removing a certain image from dataset.
BaseModel: SDXL LoRA/LyCORIS works best on the model that it was trained on.
Change notes:
v2.233->v2.233-2: Switching to a new epoch selecting strategy: reducing overfitting and improving hands.
Description
FAQ
Comments (16)
thank you for still trying to do this lora, I think your model everytime show a progress!
I think this is the first time that I can generate images that I feel like this is one from the artist's hand. For the first time, myself have trained for the same artist around 10 times, but all ends with unsatisfactory artifacts. I do not know how can you make it. Really great work. One month before I end my last try and give up, but today I see your lora, really impressive.
I'm very happy that you enjoy my work. Thank you for your support!
@silvermoong
If you don't mind, since I've failed many times in training this artist lora, I'd like to ask you specific details about training this lora.
1. did you use doujinshi (or any other image with text or mosaic)? If you used it, how did you process his tags and images?
I personally use wd14 tagger and then just crop the portrait, but because of text bubble and other comic traces, it ends up generating images with text artifacts. I don't know how you deal with this, I run into these problems when training most artist lora that have comics as their main work.
2. are there any must-do or absolutely don't-consider tips for training this kind of artist lora?
If you don't feel comfortable answering these questions, I still appreciate your willingness to share the model! This model is the first one that produces an image that feels like it was drawn by the artist himself!
@os12z regarding your questions:
1. I use doujinshi, but only full-colored ones (mostly covers). I barely use ones with multiple views.
I use Photoshop to remove texts and bubbles that are easy to remove. I give up on ones with many texts covering characters.
I keep mosaic as I can easily tag it.
I use wd14 tagger as well.
2. You may start with a better dataset to test your training parameters. (removes all images that you may have concerns and smaller dataset speeds up the tests)
You may set alpha = 4 * DIM to test even faster (higher alpha learns more quickly)
Control variables, test one parameter at a time.
Once you get satisfied training parameters, slowly adding images that you have concerns and check the outcome.
Style training is mostly the same for different artists. Train on multiple artists to see if the parameter change is improving or not.
您好，请问该如何使用这个loar，按往常的方式安装似乎没有反应
这是一个SDXL的lora，你在使用其他SDXLlora的时候有遇到这个问题吗？
@silvermoong 阿，抱歉，我之前没有使用过sdxl的lora，请问该如何使用这类lora呢?谢谢
@ni1233 基础模型需要使用SDXL类型的，你可以试试看https://civitai.com/models/260267/animagine-xl-v3
@silvermoong 谢谢解答
Hey, can I ask what settings you used to train this lora?
您好，我想问一下sdxl对于显卡是否有要求？
N卡现在可以用内存当显存使用，我测试8G显存能使用正常版的SDXL
我听说有轻量化的方法，能进一步降低显存占用。但是我没有研究过
I have one trained on Js2prony (pony based): https://civitai.com/models/320624/sdxl-artist-style-takeda-hiromitsu-for-js2prony
Pony diffusion is great. However, I don't like PonyV6 as it is yellowish. I'm testing models based on pony diffusion. Js2prony is my current favorite pony based model.
@silvermoong Thank you very much for your recommendation, I will try it right away